A Blockchain-Enabled Split Learning Framework With a Novel Client Selection Method for Collaborative Learning in Smart Healthcare

Abstract

Distributed machine learning in healthcare has great potential in training models to learn from the patients’ data distributed across different medical institutions. Recently, there has been a significant surge in research works applying federated learning(FL), a popular distributed machine learning technique in the healthcare sector. However, FL faces challenges like communication overhead and scalability to low-resource devices like Internet of Things(IoT) nodes. Addressing these issues, we propose a Blockchain-enabled split learning framework with a novel client selection algorithm for collaborative learning in healthcare. In the split learning model, the neural network is trained between the server and the clients, and the forward and backward propagation steps to update the weights happen in a collaborative way. The Blockchain platform serves the functions of decentralized model governance, decentralized identity and access management, incentive management, and client selection governance in the proposed framework. We proposed a comprehensive client selection algorithm incorporating several client features like deadline strictness, resource availability, data utility, model utility, etc. The experimental results show that the proposed split learning model achieves better results than the federated learning and cloud-centric machine learning models. Further, we also provide a hardware implementation for the proposed framework to gauge its real-world deployment feasibility.